Likelihood ratio tests are intuitively appealing. Nevertheless, a number of examples are known in which they perform very poorly. The present paper discusses a large class of situations in which this is the case, and analyzes just how intuition misleads us; it also presents an alternative approach which in these situations is optimal.

Generalized t-tests are constructed under weaker than normal conditions. In the first part of this paper we assume only the symmetry (around zero) of the error distribution (i). In the second part we assume that the error distribution is a Gaussian scale mixture (ii). The optimal (smallest) critical values can be computed from generalizations of Student's cumulative distribution function (cdf), $t_n(x)$. The cdf's of the generalized $t$-test statistics are denoted by (i) $t_n^S (x) $ and (ii) $t_n^G (x)$, resp. As the sample size $n \to \infty $ we get the counterparts of the standard normal cdf $\Phi(x)$: (i) $ \Phi^S (x):= \operatorname{lim}_{n\to \infty} t_n^S (x)$, and (ii) $ \Phi^G (x):= \operatorname{lim}_{n\to \infty} t_n^G (x)$. Explicit formulae are given for the underlying new cdf's. For example $ \Phi^G (x) = \Phi(x) $ iff $ |x| \ge \sqrt 3 $. Thus the classical 95\% confidence interval for the unknown expected value of Gaussian distributions covers the center of symmetry with at least 95\% probability for Gaussian scale mixture distributions. On the other hand, the 90\% quantile of $\Phi^G$ is $ 4\sqrt3/5 = 1.385\dots > \Phi^{-1}(0.9)=1.282\dots$.

There are many different notions of optimality even in testing a single hypothesis. In the multiple testing area, the number of possibilities is very much greater. The paper first will describe multiplicity issues that arise in tests involving a single parameter, and will describe a new optimality result in that context. Although the example given is of minimal practical importance, it illustrates the crucial dependence of optimality on the precise specification of the testing problem. The paper then will discuss the types of expanded optimality criteria that are being considered when hypotheses involve multiple parameters, will note a few new optimality results, and will give selected theoretical references relevant to optimality considerations under these expanded criteria.

Consider the problem of testing multiple null hypotheses. A classical approach to dealing with the multiplicity problem is to restrict attention to procedures that control the familywise error rate ($FWER$), the probability of even one false rejection. However, if $s$ is large, control of the $FWER$ is so stringent that the ability of a procedure which controls the $FWER$ to detect false null hypotheses is limited. Consequently, it is desirable to consider other measures of error control. We will consider methods based on control of the false discovery proportion ($FDP$) defined by the number of false rejections divided by the total number of rejections (defined to be 0 if there are no rejections). The false discovery rate proposed by Benjamini and Hochberg (1995) controls $E (FDP)$. Here, we construct methods such that, for any $\gamma$ and $\alpha$, $P \{ FDP > \gamma \} \le \alpha$. Based on $p$-values of individual tests, we consider stepdown procedures that control the $FDP$, without imposing dependence assumptions on the joint distribution of the $p$-values. A greatly improved version of a method given in Lehmann and Romano is derived and generalized to provide a means by which any sequence of nondecreasing constants can be rescaled to ensure control of the $FDP$. We also provide a stepdown procedure that controls the $FDR$ under a dependence assumption.

This research deals with massive multiple hypothesis testing. First regarding multiple tests as an estimation problem under a proper population model, an error measurement called Erroneous Rejection Ratio (ERR) is introduced and related to the False Discovery Rate (FDR). ERR is an error measurement similar in spirit to FDR, and it greatly simplifies the analytical study of error properties of multiple test procedures. Next an improved estimator of the proportion of true null hypotheses and a data adaptive significance threshold criterion are developed. Some asymptotic error properties of the significant threshold criterion is established in terms of ERR under distributional assumptions widely satisfied in recent applications. A simulation study provides clear evidence that the proposed estimator of the proportion of true null hypotheses outperforms the existing estimators of this important parameter in massive multiple tests. Both analytical and simulation studies indicate that the proposed significance threshold criterion can provide a reasonable balance between the amounts of false positive and false negative errors, thereby complementing and extending the various FDR control procedures. S-plus/R code is available from the author upon request.

After some general remarks about the interrelation between philosophical and statistical thinking, the discussion centres largely on significance tests. These are defined as the calculation of $p$-values rather than as formal procedures for ``acceptance'' and ``rejection.'' A number of types of null hypothesis are described and a principle for evidential interpretation set out governing the implications of $p$-values in the specific circumstances of each application, as contrasted with a long-run interpretation. A variety of more complicated situations are discussed in which modification of the simple $p$-value may be essential.

R. A. Fisher founded modern statistical inference in 1922 and identified its fundamental problems to be: specification, estimation and distribution. Since then the problem of statistical model specification has received scant attention in the statistics literature. The paper traces the history of statistical model specification, focusing primarily on pioneers like Fisher, Neyman, and more recently Lehmann and Cox, and attempts a synthesis of their views in the context of the Probabilistic Reduction (PR) approach. As argued by Lehmann, a major stumbling block for a general approach to statistical model specification has been the delineation of the appropriate role for substantive subject matter information. The PR approach demarcates the interrelated but complemenatry roles of substantive and statistical information summarized ab initio in the form of a structural and a statistical model, respectively. In an attempt to preserve the integrity of both sources of information, as well as to ensure the reliability of their fusing, a purely probabilistic construal of statistical models is advocated. This probabilistic construal is then used to shed light on a number of issues relating to specification, including the role of preliminary data analysis, structural vs. statistical models, model specification vs. model selection, statistical vs. substantive adequacy and model validation.

We consider concepts and models for measuring inequality in the distribution of resources with a focus on how inequality varies as a function of covariates. Lorenz introduced a device for measuring inequality in the distribution of income that indicates how much the incomes below the u$^{th}$ quantile fall short of the egalitarian situation where everyone has the same income. Gini introduced a summary measure of inequality that is the average over u of the difference between the Lorenz curve and its values in the egalitarian case. More generally, measures of inequality are useful for other response variables in addition to income, e.g. wealth, sales, dividends, taxes, market share and test scores. In this paper we show that a generalized van Zwet type dispersion ordering for distributions of positive random variables induces an ordering on the Lorenz curve, the Gini coefficient and other measures of inequality. We use this result and distributional orderings based on transformations of distributions to motivate parametric and semiparametric models whose regression coefficients measure effects of covariates on inequality. In particular, we extend a parametric Pareto regression model to a flexible semiparametric regression model and give partial likelihood estimates of the regression coefficients and a baseline distribution that can be used to construct estimates of the various conditional measures of inequality.

We consider estimation in a class of semiparametric transformation models for right–censored data. These models gained much attention in survival analysis; however, most authors consider only regression models derived from frailty distributions whose hazards are decreasing. This paper considers estimation in a more flexible class of models and proposes conditional rank M-estimators for estimation of the Euclidean component of the model

We propose a class of transformation hazard models for rightcensored failure time data. It includes the proportional hazards model (Cox) and the additive hazards model (Lin and Ying) as special cases. Due to the requirement of a nonnegative hazard function, multidimensional parameter constraints must be imposed in the model formulation. In the Bayesian paradigm, the nonlinear parameter constraint introduces many new computational challenges. We propose a prior through a conditional-marginal specification, in which the conditional distribution is univariate, and absorbs all of the nonlinear parameter constraints. The marginal part of the prior specification is free of any constraints. This class of prior distributions allows us to easily compute the full conditionals needed for Gibbs sampling, and hence implement the Markov chain Monte Carlo algorithm in a relatively straightforward fashion. Model comparison is based on the conditional predictive ordinate and the deviance information criterion. This new class of models is illustrated with a simulation study and a real dataset from a melanoma clinical trial.

In this paper, we obtain general representations for the joint distributions and copulas of arbitrary dependent random variables absolutely continuous with respect to the product of given one-dimensional marginal distributions. The characterizations obtained in the paper represent joint distributions of dependent random variables and their copulas as sums of U-statistics in independent random variables. We show that similar results also hold for expectations of arbitrary statistics in dependent random variables. As a corollary of the results, we obtain new representations for multivariate divergence measures as well as complete characterizations of important classes of dependent random variables that give, in particular, methods for constructing new copulas and modeling different dependence structures.

The results obtained in the paper provide a device for reducing the analysis of convergence in distribution of a sum of a double array of dependent random variables to the study of weak convergence for a double array of their independent copies. Weak convergence in the dependent case is implied by similar asymptotic results under independence together with convergence to zero of one of a series of dependence measures including the multivariate extension of Pearson's correlation, the relative entropy or other multivariate divergence measures. A closely related result involves conditions for convergence in distribution of $m$-dimensional statistics $h(X_t, X_{t+1}, \ldots, X_{t+m-1})$ of time series $\{X_t\}$ in terms of weak convergence of $h(\xi_t, \xi_{t+1},\ldots, \xi_{t+m-1})$, where $\{\xi_t\}$ is a sequence of independent copies of $X_t'$s, and convergence to zero of measures of intertemporal dependence in $\{X_t\}$. The tools used include new sharp estimates for the distance between the distribution function of an arbitrary statistic in dependent random variables and the distribution function of the statistic in independent copies of the random variables in terms of the measures of dependence of the random variables. Furthermore, we obtain new sharp complete decoupling moment and probability inequalities for dependent random variables in terms of their dependence characteristics.

Although regression trees were originally designed for large datasets, they can profitably be used on small datasets as well, including those from replicated or unreplicated complete factorial experiments. We show that in the latter situations, regression tree models can provide simpler and more intuitive interpretations of interaction effects as differences between conditional main effects. We present simulation results to verify that the models can yield lower prediction mean squared errors than the traditional techniques. The tree models span a wide range of sophistication, from piecewise constant to piecewise simple and multiple linear, and from least squares to Poisson and logistic regression.

Lehmann’s ideas on concepts of dependence have had a profound effect on mathematical theory of reliability. The aim of this paper is two-fold. The first is to show how the notion of a “hazard potential” can provide an explanation for the cause of dependence between life-times. The second is to propose a general framework under which two currently discussed issues in reliability and in survival analysis involving interdependent stochastic processes, can be meaningfully addressed via the notion of a hazard potential. The first issue pertains to the failure of an item in a dynamic setting under multiple interdependent risks. The second pertains to assessing an item’s life length in the presence of observable surrogates or markers. Here again the setting is dynamic and the role of the marker is akin to that of a leading indicator in multiple time series.

In the competing risks problem, an important role is played by the cumulative incidence function (CIF), whose value at time $t$ is the probability of failure by time $t$ from a particular type of failure in the presence of other risks. In some cases there are reasons to believe that the CIFs due to various types of failure are linearly ordered. El Barmi et al. studied the estimation and inference procedures under this ordering when there are only two causes of failure. In this paper we extend the results to the case of $k$ CIFs, where $k\ge3$. Although the analyses are more challenging, we show that most of the results in the 2-sample case carry over to this $k$-sample case.

In genetic studies of complex diseases, the underlying mode of inheritance is often not known. Thus, the most powerful test or other optimal procedure for one model, e.g. recessive, may be quite inefficient if another model, e.g. dominant, describes the inheritance process. Rather than choose among the procedures that are optimal for a particular model, it is preferable to see a method that has high efficiency across a family of scientifically realistic models. Statisticians well recognize that this situation is analogous to the selection of an estimator of location when the form of the underlying distribution is not known. We review how the concepts and techniques in the efficiency robustness literature that are used to obtain efficiency robust estimators and rank tests can be adapted for the analysis of genetic data. In particular, several statistics have been used to test for a genetic association between a disease and a candidate allele or marker allele from data collected in case-control studies. Each of them is optimal for a specific inheritance model and we describe and compare several robust methods. The most suitable robust test depends somewhat on the range of plausible genetic models. When little is known about the inheritance process, the maximum of the optimal statistics for the extreme models and an intermediate one is usually the preferred choice. Sometimes one can eliminate a mode of inheritance, e.g. from prior studies of family pedigrees one may know whether the disease skips generations or not. If it does, the disease is much more likely to follow a recessive model than a dominant one. In that case, a simpler linear combination of the optimal tests for the extreme models can be a robust choice.

In this paper, we determine which non-random sampling of fixed size gives the best linear predictor of the sum of a finite spatial population. We employ different multiscale superpopulation models and use the minimum mean-squared error as our optimality criterion. In multiscale superpopulation tree models, the leaves represent the units of the population, interior nodes represent partial sums of the population, and the root node represents the total sum of the population.We prove that the optimal sampling pattern varies dramatically with the correlation structure of the tree nodes. While uniform sampling is optimal for trees with “positive correlation progression”, it provides the worst possible sampling with “negative correlation progression.” As an analysis tool, we introduce and study a class of independent innovations trees that are of interest in their own right. We derive a fast water-filling algorithm to determine the optimal sampling of the leaves to estimate the root of an independent innovations tree.

We analyze the (unconditional) distribution of a linear predictor that is constructed after a data-driven model selection step in a linear regression model. First, we derive the exact finite-sample cumulative distribution function (cdf) of the linear predictor, and a simple approximation to this (complicated) cdf. We then analyze the large-sample limit behavior of these cdfs, in the fixed-parameter case and under local alternatives.

Local asymptotic minimax risk bounds in a locally asymptotically mixture of normal family of distributions have been investigated under asymmetric loss functions and the asymptotic distribution of the optimal estimator that attains the bound has been obtained.

This paper concerns estimating a probability density function $f$ based on iid observations from $g(x) = W^{-1} \, w(x) \, f(x) $, where the weight function $w$ and the total weight $W = \int \, w(x) \, f(x) \, d x $ may not be known. The length-biased and excess life distribution models are considered. The asymptotic normality and the rate of convergence in mean squared error (MSE) of the estimators are studied.

We establish consistency and asymptotic normality of the minimum density power divergence estimator under regularity conditions different from those originally provided by Basu et al.